On Study Of The Feature Extraction Of Human Behavior

In recent years,with the rise of artificial intelligence,the application requirements of computer vision technology have been continuously improved,and research on human behavior recognition is also very important.Human behavior recognition has broad application prospects in intelligent monitoring,human-computer interaction,video retrieval,annotation,and motion analysis.Therefore,the recognition of human behavior has been received more and more attention.This paper extracts the local characteristics of the behavior of the skeletal node from the skeletal sequence data.The vector of locally aggregated descriptors is used to aggregate the local features into coarse classification feature vectors and extract the features again to improve the feature information discrimination.The experiments show the effectiveness of the proposed algorithm.The main work of this paper includes the following aspects:First of all,a unified description of bone data is carried out.Since the skeleton representations of the three datasets used in this paper are not uniform,the MSR-Action 3D dataset and the UT Kinect-Action 3D dataset are represented by 20 nodes,and the Florence 3D Actions datasets is represented by 15 nodes.In order to prepare for the extraction of local features in the next stage,this paper use the NITE model to extract bone data,unify the bones into a 15-node representation,and filter the bone data using the mean filter.Then,this paper extracts the local behavior characteristics of the processed bones.In this paper,divides the 15 joints into several groups according to the principle from the middle to the extremities.Each group is used as a local feature to extract the relative displacement characteristics and relative position of the skeleton sequence,then combine the two features as local features.Then it uses the vector of locally aggregated descriptors(VLAD)to represent local features as a dimension-specific feature vector.Finally,this paper uses the large margin nearest neighbor(LMNN)algorithm to extract the quadratic features.It uses the stochastic gradient descent algorithm to globally find the metric matrix.The original features multiplied by the metric matrix,not only effectively reduce the dimension of the feature,but also improve the discriminability of the feature information.Finally,using the k-nearest neighbor(K-NN)and support vector machine(SVM)classification algorithms,we have done experimental analysis in MSR-Action 3D dataset,UT Kinect-Action dataset and Florence 3D Actions datasets.Experiments show that the recognition rate is increased by about 5% through the secondary extraction of behavioral features.